Enhancing Human Activity Recognition in Smart Homes with Self-Supervised Learning and Self-Attention.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Sensor Data Analysis

Background:

• Deep learning models are crucial for human activity recognition (HAR) using ambient sensors, especially for telemonitoring older adults.
• Collecting large, annotated sensor datasets is challenging due to time and cost constraints of manual annotation.
• Existing methods struggle with limited or no annotated data in real-world HAR applications.

Purpose of the Study:

• To propose a novel self-supervised learning model, AttCLHAR, for HAR using ambient sensor data with minimal or no annotations.
• To leverage the SimCLR framework with a self-attention mechanism and sharpness-aware minimization (SAM) for improved feature learning.
• To enhance the model's ability to capture spatial and temporal dependencies in sensor data for telemonitoring applications.

Main Methods:

• Developed AttCLHAR, a model with unsupervised pre-training and fine-tuning phases using a shared encoder (CNNs and LSTM).
• Integrated a self-attention layer to focus on relevant input sequence segments and SAM to improve generalization.
• Conducted extensive evaluations on three CASAS smart home datasets (Aruba-1, Aruba-2, Milan).

Main Results:

• AttCLHAR demonstrated superior performance compared to baseline SimCLR, SimCLR with SAM, and SimCLR with self-attention.
• The model achieved significant advancements, particularly in semi-supervised and transfer learning settings.
• Experimental results confirmed the effectiveness of self-supervised learning in extracting insights from unlabeled ambient sensor data.

Conclusions:

• AttCLHAR offers a robust solution for human activity recognition in scenarios with limited annotated data.
• The integration of self-attention and SAM enhances the model's ability to learn from unlabeled sensor data.
• This approach represents a significant step forward in leveraging ambient sensors for telemonitoring and HAR.